Proposal of a Long - Text Generation Method Considering Article Structure Using ChatGPT

This study presents an innovative approach for the controlled generation of long Japanese texts using Large Lan-guage Models (LLMs). By incorporating abstracted structural el-ements from article data, we aim to maintain text length, content relevance, and coherence. Our method employs two LLMs: one to extract structural elements and another to generate texts based on these elements. Evaluations against baseline models using the Livedoor news corpus indicate that our approach outperforms in producing controlled-length texts with enhanced vocabulary and expression diversity. LLM-based evaluations and LIME analysis reveal that our method generates more original and content-relevant texts while adhering to the desired structures. However, the effectiveness varies across categories, suggesting the need for refined control mechanisms. This research offers valuable insights into controlled long-text generation in Japanese, with implications for various NLP tasks. Future work will focus on developing more sophisticated control methods and exploring applications in other languages and domains.